[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]
Re: [PATCH 2/2] target/riscv: fence: reconcile with specification
From: |
Philipp Tomsich |
Subject: |
Re: [PATCH 2/2] target/riscv: fence: reconcile with specification |
Date: |
Fri, 12 Aug 2022 16:17:00 +0200 |
Happy to lower it back into the decode file.
However, I initially pulled it up into the trans-function to more
closely match the ISA specification: there is only one FENCE
instruction with 3 arguments (FM, PRED, and SUCC).
One might argue that the decode table for "RV32I Base Instruction Set"
in the specification lists FENCE.TSO as a separate instruction, but
the normative text doesn't (and FENCE overlaps FENCE.TSO in the
tabular representation) — so I would consider the table as
informative.
I'll wait until we see what consensus emerges from the discussion.
Philipp.
On Fri, 12 Aug 2022 at 15:21, Peter Maydell <peter.maydell@linaro.org> wrote:
>
> On Fri, 12 Aug 2022 at 14:17, Philipp Tomsich <philipp.tomsich@vrull.eu>
> wrote:
> >
> > Our decoding of fence-instructions is problematic in respect to the
> > RISC-V ISA specification:
> > - rs and rd are ignored, but need to be 0
> > - fm is ignored
> >
> > This change adjusts the decode pattern to enfore rs and rd being 0,
> > and validates the fm-field (together with pred/succ for FENCE.TSO) to
> > determine whether a reserved instruction is specified.
> >
> > While the specification allows UNSPECIFIED behaviour for reserved
> > instructions, we now always raise an illegal instruction exception.
> >
> > Signed-off-by: Philipp Tomsich <philipp.tomsich@vrull.eu>
> >
> > ---
> >
> > target/riscv/insn32.decode | 2 +-
> > target/riscv/insn_trans/trans_rvi.c.inc | 19 ++++++++++++++++++-
> > 2 files changed, 19 insertions(+), 2 deletions(-)
> >
> > diff --git a/target/riscv/insn32.decode b/target/riscv/insn32.decode
> > index 089128c3dc..4e53df1b62 100644
> > --- a/target/riscv/insn32.decode
> > +++ b/target/riscv/insn32.decode
> > @@ -150,7 +150,7 @@ srl 0000000 ..... ..... 101 ..... 0110011 @r
> > sra 0100000 ..... ..... 101 ..... 0110011 @r
> > or 0000000 ..... ..... 110 ..... 0110011 @r
> > and 0000000 ..... ..... 111 ..... 0110011 @r
> > -fence ---- pred:4 succ:4 ----- 000 ----- 0001111
> > +fence fm:4 pred:4 succ:4 00000 000 00000 0001111
> > fence_i 000000000000 00000 001 00000 0001111
> > csrrw ............ ..... 001 ..... 1110011 @csr
> > csrrs ............ ..... 010 ..... 1110011 @csr
> > diff --git a/target/riscv/insn_trans/trans_rvi.c.inc
> > b/target/riscv/insn_trans/trans_rvi.c.inc
> > index ca8e3d1ea1..515bb3b22a 100644
> > --- a/target/riscv/insn_trans/trans_rvi.c.inc
> > +++ b/target/riscv/insn_trans/trans_rvi.c.inc
> > @@ -795,7 +795,24 @@ static bool trans_srad(DisasContext *ctx, arg_srad *a)
> >
> > static bool trans_fence(DisasContext *ctx, arg_fence *a)
> > {
> > - /* FENCE is a full memory barrier. */
> > + switch (a->fm) {
> > + case 0b0000:
> > + /* normal fence */
> > + break;
> > +
> > + case 0b0001:
> > + /* FENCE.TSO requires PRED and SUCC to be RW */
> > + if (a->pred != 0xb0011 || a->succ != 0b0011) {
> > + return false;
> > + }
> > + break;
> > +
> > + default:
> > + /* reserved for future use */
> > + return false;
> > + }
>
> I think it would be neater to do this decode in the
> .decode file, rather than by hand in the trans function.
>
> thanks
> -- PMM